Title: Forecast evaluation and dynamic panels
Authors: Charles Saunders - University of Western Ontario (Canada) [presenting]
Abstract: Calculation of forecast evaluation statistics for time series models requires a large number of predicted observations for appropriate inference when comparing models. Panel data can be limited to few time periods, so dynamic panel estimation methods correct for the well-known incidental parameter bias problem. A panel Diebold-Mariano statistic is constructed that includes the cross-sections of the panel. This allows for valid forecast inference with a minimum of a single cross-section of predicted errors. The panel Diebold-Mariano statistic is examined via simulation for critical values for both stationary and nonstationary models, finding that the standard normal critical values are appropriate for stationary panels with Large-$N$. The simulation study examines the effect of using a consistent (moment-based) or an inconsistent (fixed-effects) estimator. Forecast evaluation of nonstationary panels lead to size distortions for some GMM estimators, that are known to perform poorly at the unit boundary.